Meeting minutes
Decision tree
<julierawe> https://
Julie: The link above is to the main draft.
<julierawe> https://
Julie: The second link is to the decision tree section of the draft - there are three versions in this section.
Julie: In July - we had a broader take on the decision tree. I am hoping to focus our meeting today on what Rachael has suggested.
Julie: We got feedback that what we have written is very focused on what author's should do, but we need to look at what tools can do.
Julie: Right now, it's very short. Can the content be rendered and does it contain non-literal language?
Laura: My question was, "What would make it unavailable to user agents.
Jan: Assistive technologies are also user agents.
<julierawe> W3C definition of user agent: https://
<julierawe> Jan Will the inclusive piece get missed?
<julierawe> Jan It concerns me that there is now no emphasis on authors
<julierawe> Jan It should be a combination of authors and user agents
JohnR: provided a link to the W3C user agent definition.
Len: As an example, a voice UI makes the experience different than other UIs.
Laura: We have to know for sure that user agents can handle it, and I don't think that first question does that,
Julie: "Can user agents access it?" is the first question, but then we might be getting into the quality of it.
JohnR: If a user agent can't do it, then a human has to do it.
<julierawe> Jan If we get to this level of simplicity, we are doing a disservice to people with multiple disabilities
<julierawe> Jan The tools may not be designed to meet the nuanced problems of cognitive accessibility
<julierawe> Jan We're supposed to expect that the authors care enough to educate themselves and design inclusively
<julierawe> Jan I think we lose that if we think AI is going to solve all of our problems
<julierawe> Jan If we dilute the responsibility of humans, then we're not doing our jobs
Julie: When we spoke to Mike Gower, we talked about the limitation of tools. The hope is that some portion of the work can be done by user agents, but that humans will have to step in when the tools are not sufficient.
JohnR: Mike was not really focusing on AI, he was focusing on tools that can identify metaphors and that there are tools like that that can be used.
Julie: If you come back to the quality question and you assume there are tools that can do this work, for example - recognize metaphors, but if a tool fails to notice a metaphor, then the author has to look. You run a quick check and then you have to take a look yourself. In the simplified version of the decision tree, it doesn't necessarily
identify responsibility - human or tool.
Julie: Is there something in the simplicity that gives flexibility, or does it let authors off the hook?
Len: Ultimately you're responsible for your site being understood. Maybe that should be said up front. At the end of the day, the tool is not hurting or harming people - it's the decisions made by people in the flow of designing the content.
Julie: Is it possible that in Rachael's version, there is still responsibility for authors? Is the simplicity helping, or letting authors off the hook. How would we do it differently?
JohnK: Is non-literal language the only implied meeting? What about jargon?
Julie: I think that jargon gets covered by something else than implied meaning.
JohnR: We could have exceptions - like industry terms.
JohnR: We're being required to address every language in the world, but these tools almost exclusively focus on English. That is where the industry has always been and I am quite concerned about that.
Julie: It may be that we start with focusing on English and write this in a way that it leaves room for tools to evolve in other languages. I think JohnK was looking at a study about a particular implied meaning in Korean. Google Translate and other translation software applications are probably looking at these kinds of issues.
JohnR: I don't know what the relevance is to translation tools.
Julie: If I were to translate an English phrase that included an idiom and then back translated that phrase, would it translate it correctly?
JohnR: I do not believe that would work because existing translation applications don't have the ability to translate non-literal language.
Julie: I guess this raises the question of "What do user agents do with the content." The first question in Rachael's decision tree might mean that we have to have a lot of faith in user agents.
Julie, JohnK, JohnR: Translated the phrase, "Getting my driver's license was a piece of cake" into Spanish and French and the translation software translated the phrase into idioms that are used in other languages. It appears that the tools are recognizing non-literal language.
Len: We might be assuming that users are using their primary devices. Users can turn certain things on in their browsers, but then might use a different device. How concerned should we be about that? Sometimes tools might not be able to recognize non-literal languages until they are trained.
Len: I think we're going to need more development in the tools before we can trust them.
Julie: Coming back to Rachael's decision tree, could we add another question "If user agents can access the text, does it flag any phrases it does not know?" and if so, then the author needs to deal with it.
JohnR: There are tools where you can tell it to flag metaphors, similes, etc., but they are for English and not for multiple languages.
Julie: Len, you mentioned a tool you're working on includes Russian - we could look to see if there is anything in Hindi, Mandarin, and Arabic that will address non-literal language.
Julie: translation packages have to understand the non-literal language in order to translate it, so I think there's hope that the tools might be able to eventually get better.
<kirkwood> “Google Translate does not attempt to understand the full meaning of sentences as a human translator would, and therefore does not handle implied meaning. Instead, it uses statistical patterns and word-clustering algorithms to translate text”
Len: where I work, we have great writers and a large team that can work on this, but I worry about smaller businesses - how do we write guidance that is doable by them.
Julie: Can we do some investigation this week on what tools can do?
JohnK: I was curious with how the translation software we looked at were able to pull out implied meaning - there is something in there.
Len: Those are specifically designed for certain contexts - for example, conversational language, as opposed to written languag for a web site.
JohnR: As JohnK was doing his checking, I was checking other languages and found some research in some cases.
JohnK: Are we good on our own definition of implied meaning?
Julie: We need to have a clear defintion of what we mean by non-literal language.
JohnK: Yes, so that we can build on it.
JohnR: We have to account for whether the tools themselves are digitally accessible to people with disabilities.
Julie: So we have 2 questions, Do the tools exist, and if so, are they accessible to people with disabilities who may need to use them?
Julie: We also have questions about whether the tools exist in other lanugages.
Julie: Please spend some time this week, looking around for tools.